Bearing Faults Classification Using THH and Neural Network

Transcription

1 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 9, Issue 2, No 1, March 2012 ISSN (Onlne): Bearng Faults Classfcaton Usng THH and Neural Network A. KABLA and K. MOKRANI Department of Electrcal Engneerng, Unversty of Bejaa, 06000, Bejaa Algera Abstract The nducton machne has many advantages: ts specfc power, ts strength, relatvely low manufacturng cost and mnmal mantenance. But despte all these qualtes, a number of faults can affect the lfe of the machne showng premature falures. The purpose of preventve mantenance n real tme, we ntroduce a new sgnal processng technque based on Hlbert- Huang Transform (HHT) and margnal spectrum. Frstly, the current sgnals are decomposed nto several ntrnsc mode functon (IMFs) usng the emprcal mode decomposton (EMD). The Hlbert Huang spectrum for each IMF s an energy representaton n the tme-frequency doman usng the nstantaneous frequency. The margnal spectrum of each IMF can then be obtaned. The next step s the classfcaton of faults detected by the applcaton of neural network on IMFs. Tests on real sgnals show that the margnal spectrum of the second IMFs can be used for the detecton and classfcaton of bearng faults. The proposed approach provdes a vable sgnal processng tool for an onlne machne health status montorng. Keywords: Sgnal processng, bearng faults, Hlbert-Huang transform, emprcal mode decomposton, neural networks. 1. Introducton The asynchronous machne occupes an mportant place n the ndustry. It s used n many dfferent applcatons: wnd, mltary, electrc drve for hgh speed tran and pumpng. To ensure contnuty of operaton, establshment of mantenance programs are requred. Tradtonally the mantenance procedure, known as remedy, was to repar or replace faulty equpments. A new approach, called predctve mantenance, s the detecton and localzaton of faults and falures and act earler to mnmze ther secondary effects [1]. Studes on the analyss of the current sgnal (MCSA: Motor Current Sgnature Analyss) show that n addton to the nformaton contaned n the vbratons, nformaton specfc to electrcal phenomena appears n the stator current sgnal [2]. Many methods based on the MCSA have been developed. These methods nclude the assessment of the power spectrum, the fast Fourer transform (FFT), spectrum analyss of the envelope. It turned out that they are effectve n the detecton of bearng faults. However, they are lmted to statonary sgnals. To treat non-statonary sgnals, several tme-frequency analyss tools are commonly used such as Fourer transform short-term (STFT) [3], the Wgner-Vlle Dstrbuton (WVD) [4], and RTF (Tme-frequency representaton) of Cohen's class and the wavelet transform (WT) [5]. The man drawback of these methods s that they depend on dfferent parameters. For example, selecton of a sutable wndow sze s requred when applyng the STFT to match wth the specfc frequency content of the sgnal, whch s generally not known a pror. Wavelets requre the specfcaton of a core or a core functon. However, there s no unversal core. In addton, one lmtaton of TFR, such as the WVD s the presence of nterferng terms whch affects the nterpretaton and readablty of the resultng representatons [6]. The tme-frequency smoothng can reduce nterferences but t ntroduces tme and frequency localzaton errors. In ths paper, we ntroduce a new approach for analyzng non- statonary sgnals developed by Huang et al [7]. Frst the sgnal s decomposed nto several ntrnsc mode functon (IMFs) usng the emprcal mode decomposton (EMD). Then the Hlbert transform s appled to each IMF. Therefore, a tme-frequency dstrbuton s obtaned. The Hlbert-Huang Transform s appled n ths paper for detecton of bearng faults. After detecton, you go to the classfcaton of these faults s a very mportant operaton for possble nterventon n preventve mantenance. 2. Emprcal Mode decomposton The EMD s defned by a sftng process. It can decompose a mult-components sgnal nto a seres of IMFs. Huang et al [7] have defned the IMFs as a functon class that satsfes two condtons: 1) At any pont, the mean value between the envelope defned by local maxma and the envelope defned by the local mnma s zero. 2) The number of extrema and the number of zerocrossngs are ether equal to each other or dffer by at most one. Copyrght (c) 2012 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

2 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 9, Issue 2, No 1, March 2012 ISSN (Onlne): To extract the IMFs, the sftng process used s defned n [7]. Havng obtaned the IMFs, we apply the Hlbert transform to each IMF. To calculate the nstantaneous characterstcs (frequency and ampltude) of each IMF, the analytc sgnal t assocated to where: c t z s used: t c t j Hc t z (1) 1 c H c t P t z t can be expressed by: z t a t expjw t d (2) (3) The ampltude and nstantaneous phase are defned by: a 2 2 t c t H c t (4) c t t H t arctan (5) c The nstantaneous frequency of z t, s smply the dervatve of the nstantaneous phase: d t (6) dt Thus, the orgnal sgnal can be expressed as : x n t Re a t exp j w t 1 Where the resdue r n was omtted. real part of a complex quantty. dt (7) Re Denotes the Ths tme-frequency dstrbuton s desgnated as the H w,t : Hlbert-Huang spectrum H n w,t Re a t exp j w t 1 Equaton (8) allows us to represent the nstantaneous ampltude and frequency n three dmensons, n whch the ampltude s the heght n the tme-frequency plane. The tme ntegral of Huang-Hlbert spectrum s the margnal Hlbert spectrum h w defned as: T hw H w,tdt (9) where T s the sgnal duraton. 0 dt (8) The margnal spectrum offers a measure of the energy at each frequency. It represents the cumulated ampltude over the entre data span n a probablstc sense. Therefore, the margnal spectrum of each IMF can be defned, as: T h w H w,tdt (10) 3. Descrpton 3.1 Prncpal faults n the machne 0 A study conducted for IEEE [9], establshed a statstcal flaws that may occur on asynchronous machnes; Bearng: 41%, Stator: 37%, Rotor: 10%, and other: 12%. Ths dstrbuton shows that faults come manly from bearngs. Faulty bearngs causes ar gap eccentrctes due to rregular moton of the rotor. Theses eccentrctes affect the stator current due to the varatons of the electromagnetc feld. 3.2 Characterzaton of ball bearngs faults A ball bearng fault s characterzed by a contnual repetton of faulty contacts wth the bearng outer and nner cage. And nsofar as the ball bearng supports the rotor, each fault wll produce a radal moton of the rotor relatve to the stator [10]-[11]. Accordng to R.Schoen [12], these varatons generate stator currents at frequences: f k. f where : k = 1, 2, 3, (11) a fault Wth: f a ; power source frequency f fault ; characterstc frequency nduced by the fault. The characterstc frequences of the faults depend on the bearng dmensons [2], [12]. The bearng faults can be classfed as nner rng, outer rng or ball. 3.3 Descrpton of montorng system The tests were carred on the test bed at the LEG Laboratory of Grenoble. The defects were created artfcally by an electrcal eroson of a 1 mm dameter hole at the outer rng, nner rng or ball [13]. The montorng system of the stator current used s composed of: - A three-phase power source wth a frequency of 51Hz. Copyrght (c) 2012 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

3 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 9, Issue 2, No 1, March 2012 ISSN (Onlne): A sampler whch am to acqure the three phase voltages and currents of the stator. After a low pass flterng, sgnal are sampled at 10 khz. - An asynchronous squrrel cage wth the followng characterstcs: Table 1: Characterstcs of the Machne Parameter Power Power Frequency Motor connecton Phase voltage Rated speed Number of pole par Number of rotor slots Number of stator slots Value 1.1KW 51 Hz Y 400V 1445tr/mn P= Fg. 2 Margnal spectrum of IMF Motor wth outer race fault 4. Characterzaton of bearng fault by HHT The purpose of ths secton s to descrbe the sgnature bearng fault as t has been observed on the steady state sgnals. Analyses were performed on blocks of ponts; the 10 KHz samplng frequency allows a spectral resoluton of 1 6 Hz. 4.1 Healthy motor After havng decomposed the healthy sgnal nto IMF the margnal spectrum to IMF 2 and IMF 3 are shown n the followng fgures: The margnal spectrum of IMF 2 and IMF 3 are shown n the followng fgures: For a rotaton speed of 1445 rpm, the theoretcal characterstc frequency of ths fault s: f out = 76.3 Hz. theoretcally t generates, n the stator current spectrum, lnes at frequences shown n table 2. k Table 2: Theoretcal frequences of outer race fault f fault fa k. fout f fault f a k. f out The margnal spectrum of the IMF 2 and IMF 3 are shown n the followng fgures: *fa Ampltude fa-2fo 1000 Fg. 1 Margnal spectrum of IMF Fréquence(Hz) Fg. 3 Margnal spectrum of IMF 2 Copyrght (c) 2012 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

4 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 9, Issue 2, No 1, March 2012 ISSN (Onlne): The margnal spectrum of the IMF 2 and IMF 3 are shown n the followng fgures: Frequency Fg. 4 Margnal spectrum of IMF fa-frot 0.35 Fg. 6 Margnal spectrum of IMF 2 Am pltude fa+frot Fg. 5 Margnal spectrum of IMF3 for dfferent frequency range. 4.3 Motor wth nner race fault Fg. 7 Margnal spectrum of IMF2 for dfferent frequency range The cnematc frequency of ths fault s f n =124.3Hz. Theoretcally, an nner race fault generates spectrum lnes at frequences as shown n table 3. Table 3: Theoretcal frequences of nner race fault k f fault fa k. fn f fault fa k. fn Fg. 8 Margnal spectrum of IMF 3 Copyrght (c) 2012 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

5 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 9, Issue 2, No 1, March 2012 ISSN (Onlne): Motor wth ball fault The cnematc frequency of ths fault s f b = 99.6Hz, theoretcally t generates, n the stator current spectrum, lnes at frequences as shown below k Table 4: Theoretcal frequences of ball fault f fault fa k. fb f fault fa k. fb The margnal spectrum of the IMF2 and IMF 3 are shown n the followng fgures: Fg. 11 Margnal spectrum of IMF2 For dfferent frequency range Fg. 9 Margnal spectrum of IMF 2 Fg. 12 Margnal spectrum of IMF2 for dfferent frequency range Fg. 10 Margnal spectrum of IMF2 for dfferent frequency range Fg. 13 Margnal spectrum of IMF Results From the tests results, we can say that: Copyrght (c) 2012 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

6 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 9, Issue 2, No 1, March 2012 ISSN (Onlne): bearng faults affect the margnal spectrum of the IMF 2 and 3 of the stator current, the ball fault s easer to detect, Outer race Inner race ball faults are characterzed by frequences correspondng to the theoretcal outer and nner rng faults. The ball faults can be consdered as an outer and nner rng faults [9], the ampltude of the ray correspondng to faults s very small compared to the harmoncs of the stator current, ths could be due to the fact that the faults are not severe enough to generate large ampltude streaks. Ball Health 5. Implementaton of the classfcaton by Neural Network Accordng to expresson (11), a fault creates a frequency shft of the stator current spectrum; ths shft s proportonal to the characterstc frequency of the fault. We have shown n the prevous secton that the margnal spectrum of IMF 2 characterzes the bearng fault. We propose to analyze the IMF 2 for the extracton of dscrmnatve features of the faults. The sgnals of IMFs 2 were represented by ther AR models; we estmate the order p and coeffcents (a) of each sgnal by mnmzng the fnal predcton error (FPE). To make an automatc classfcaton of the fault we chose the neural network MLP (Mult-layer Perceptron) because t seems sutable for classfcaton problems [14], [15], accordng to the partton approach. Order p = 4 was set for the parameterzaton of the test database, so the sze of the nput of each network s N E =4. The estmated weght for each learns was carred out usng standard technque of back-propagaton gradent. To operate ths classfer, we adopted a smple decson rule: use of the maxmum output found that the class presents. 6. Result of the classfcaton The classfer contan four sub-networks of hdden neurons and one output neuron each, whch are respectvely dedcated to the outer race fault class, nner race fault class, ball fault class and healthy class. After learnng from each sub-network (wth 400 sgnals), we performed the test on 80 sgnals (20 healthy, 20 nner race fault, 20 outer race fault and 20 ball fault). The followng fgure shows the output of each sub-network accordng to the test sgnals. Fg.14 Evaluaton of the output of each neural network based on test sgnals After applyng the decson rule we obtaned classfcaton rates as follows: Table 5: Classfcaton of Fault Rate Outer Inner Health Ball race race y Outer race 95% 5% 0% 0% Inner race 0% 100% 0% 0% Ball 5% 0% 95% 0% Healthy 0% 0% 0% 100% Average classfcaton rate = 98.75% 6.1. Interpretaton The classfcaton by the partton approach allowed us to obtan very good result. Thus the classfcaton rate obtaned s hgher than that gven n [16] whch use the AR model of the envelope of the stator current and neural networks. 7. Concluson In ths work, to overcome the lmtatons of tradtonal tme/frequency analyss methods, we appled a new method for the detecton of bearng faults n electrc machnes. The HHT and t margnal spectrum have proven Copyrght (c) 2012 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.

7 IJCSI Internatonal Journal of Computer Scence Issues, Vol. 9, Issue 2, No 1, March 2012 ISSN (Onlne): ther effcency for characterzng bearng faults. We can say that the analyss of the stator current usng EMD s a promsng method whch can allows an onlne montorng of the electrc machne status. We presented an applcaton of the classfcaton of bearng faults usng the AR model to the IMF 2 of stator current and neural networks, whch have yelded very nterestng classfcaton performance, especally n real tme. References [1] G. Rostang, "Dagnostc de Défaut dans les Entraînements Electrques ", Thess. Grenoble Insttute of Technologe, [2] F.Tafnne, K.Mokran, K.Hamasse, "Dagnostc des Machnes Asynchrones par l Analyse Spectrale du Courant Statorque," In Proc. Natonal Conference Electrcal Engneerng, (Unversty of Taret), November 2004, pp [3] Y. Zhang. Hlbert-Huang transform and margnal spectrum for detecton of bearng localzed defects ", IEEE Proceedng of the 6 th World Congress on Intellgent Control and Automaton, June 21-23, 2006, pp [4] W.J.Staszewsk, K. Worden, G.R.Tomlnson, "The-frequency analyss n gearbox fault detecton usng the Wgner-Vlle dstrbuton and pattern recognton", Mechancal Systems and Sgnal Processng, Vol.11, n.5, pp , [5] H. Sharabaty. " Dagnostc de somnolence d un opérateur: analyse automatque de sgnaux physologques», Thess, Unversty of Toulouse, [6] J-C. Cexus. "Analyse des sgnaux non-statonnare par transformaton de Huang, opérateur de Teager-Kaser, et transformaton de Huang- Teager (THT) ", Thess, Unversté of Rennes 1, December [7] N.E. Huang, Z. Shen, S.R. Long, M.C. Wu, H.H. Shh, Q. Zheng, N.C. Yen, C.C. Tung, H.H. Lu. "The emprcal mode decomposton and the Hlbert spectrum for nonlnear end non-statonary tme seres analyss", Proceedng of the Royal Socety of London, 454: , 1998.M. Young, The Techncal Wrters Handbook. Mll Valley, CA: Unversty Scence, [8] Q. He, H.J. Zhang, H.L. Zhou. "Evaluaton of Stress Usng Ultrasonc Technque Based on Hlbert-Huang Transform". Internatonal Symposum on Instrumentaton Scence and Technology, Conference Seres 48, pp , [9] IEEE Motor relablty workng group, "Report on large motor relablty survey of ndustral and commercal nstallatons, " IEEE Trans. on Industry applcaton, vol. IA-21, no. 4, July-August 1985, pp [10] H.Razk, "Le contenu spectral du courant absorbé par la machne asynchrone en cas de défallance, un état de l art ", Revew 3 EI, n 29, June 2002, pp [11] B.Rason, " Détecton et localsaton de défallances sur un entraînement électrque ", Thess, (INPG) Grenoble, September [12] R.R.Schoen, T.Habetler, F.Kamran, R.Bartheld, «Motor Bearng Damage Detecton Usng Stator Current Montorng,", IEEE Trans. on Industry Applcatons. vol. 31, no. 6, Nov [13] O.Butscher. " Dagnostc de la Machne Asynchrone ". Electrcal engneerng, (INPG) Grenoble, September [14] B.L, M.Chow, Y.Tpsuwan, J.Fellow, "Neural-Network-Based motor rollng bearng fault dagnoss, " IEEE Trans. on Industral Electroncs, vol.47, no.5, Oct [15] G.Dreffus, J.M.Martnez, M.Samueldes, M.B.Gordon, F.Badran, S.Thra, L.Herault, Réseaux de Neurones : Méthodologe et Applcatons, Edton Eyrolles, Pars, [16] F.Tafnne, K.Mokran, J.Anton, "Influence de la charge sur le dagnostc des machnes asynchrones" n Second Internatonal Conference on Electrcal Systems ICES 06, May , Oum El Bouagh, Algera. Copyrght (c) 2012 Internatonal Journal of Computer Scence Issues. All Rghts Reserved.